https://github.com/themcaffee/road-rage

Use deep learning to set red light timings to optimize throughput
https://github.com/themcaffee/road-rage

Last synced: 26 days ago
JSON representation

Use deep learning to set red light timings to optimize throughput

• Host: GitHub
• URL: https://github.com/themcaffee/road-rage
• Owner: themcaffee
• License: mit
• Created: 2018-09-09T01:54:55.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2023-02-15T21:32:12.000Z (almost 2 years ago)
• Last Synced: 2023-09-09T16:46:11.057Z (over 1 year ago)
• Language: Python
• Size: 85 KB
• Stars: 1
• Watchers: 1
• Forks: 0
• Open Issues: 3
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Road Rage

Hate traffic? Me too. Traffic lights should be smarter and controlling

them is a juicy optimization problem. This experiment uses reinforcement

learning to dynamically control stop lights to optimize getting people

where they want to go.

## Usage

```

# Install sumo the traffic simulator

sudo apt-get install sumo sumo-tools sumo-doc

export $SUMO_HOME=/usr/share/sumo

# Setup and install requirements

python3 -m virtualenv venv

source venv/bin/activate

pip install -r requirements.txt

# Run training and evaluation

python run.py

```

## Options

```

python run.py --help

Usage: run.py [options]

Options:

  -h, --help            show this help message and exit

  --gui                 Run the GUI version of sumo

  --type=TYPE           The type of prediction to use

  --training-steps=TRAINING_STEPS

                        The number of simulation steps to train for

  --training-max-steps=TRAINING_MAX_STEPS

                        The maximum number of steps during training per

                        episode

  --training-warmup=TRAINING_WARMUP

                        Steps to take randomly before prediction

  --eval-episodes=EVAL_EPISODES

                        Number of episodes to evaluate for

  --eval-max-steps=EVAL_MAX_STEPS

                        Max simulation steps per episode during training

```

## Results

Reward (higher is better):

- DQN (trained 200000 epochs): 13027

- random: 6965

- timed: 1495